andreakeating1982/image-generation

Image generation

A good image does its job: it sells, explains, identifies or decorates in the place it will be used. Decide the job and the route before writing a prompt.

Pick the route

The image must...	Route
Plot data with real axes, scales or numbers	Python chart (see the data-analysis skill). Never an AI image
Show exact facts: times, prices, schedules, itineraries, timetables, comparison or spec tables, checklists	HTML layout rendered to PNG: references/static-layouts.md
Show a structure: flowchart, sequence, org chart, architecture, mind map	Mermaid rendered to PNG: references/static-layouts.md. Make a polished AI version afterwards only if asked
Be a visual asset: poster, ad, social post, thumbnail, illustration, product shot, character, logo concept, icon, UI mockup image, infographic with a few key facts	AI image: media_generate_image
Change an existing image: edit, restyle, swap the background, add or remove objects, translate its text, combine people or products, upscale or restore	AI edit: media_generate_from_images
Resize, crop, convert, compress or stitch with no new content	Pillow or ffmpeg in the sandbox
Be a working web page, app or prototype	Build the software. An image is the wrong deliverable

For real people, places, products or events the user wants to see as they are, find real photos with the search tool (type image). Do not generate them.

Tools

• media_generate_image(prompt, orientation) generates with the image model the user selected (GPT Image by default, which renders text well). orientation is square, landscape or portrait.
• media_generate_from_images(prompt, image_urls, orientation) edits or combines images. image_urls takes URLs or sandbox paths such as /home/user/uploads/photo.jpg. Pass several images to combine them ("put the person from the first image into the scene of the second").
• Reading this skill unlocks the media tools. If they are still missing, call media_generate with a one-line brief.
• Each image uses the user's image credits. When the plan limit is reached, say so and suggest upgrading. Never fall back to drawing the image with Python.

Brief, then prompt

Settle these before prompting. Ask only for what you cannot sensibly assume (usually exact text, brand colours or reference photos):

• Purpose: what the image must achieve, and where it goes (feed post, website hero, print, slide, app store).
• Format: orientation and safe areas (text away from edges, room for a headline).
• Subject: the specific person, product, scene or concept.
• Style: photographic, illustration, 3D, flat vector, brand style, or matching a reference.
• Text: the exact words, organised as headline, subhead, labels and call to action, or "no text".
• Must keep: identity, product shape, logo, colours, everything outside the edit.

Prompt patterns for new images, edits and each scenario are in references/prompt-recipes.md. Read it for anything beyond a simple single image.

Rules

• Text in images. Put the exact words in quotes inside the prompt and let the model render them. Keep it to a headline and a few short lines. Never generate a blank background and then type text over it with Python. For dense exact information, use the HTML layout route instead.
• Consistency. For a recurring character or product, make one master image, then derive other poses and scenes from it with media_generate_from_images.
• Sets. For icon packs, card decks, product variants and poster series, make one image per item with the same style sentence. Do not combine them into one sheet unless asked.
• Edits keep everything else. Say exactly what changes and what must stay the same (face, pose, product, background, lighting, colours).
• Transparent assets. The tools return opaque images. Generate on a plain solid background that does not appear in the subject, then remove it: a colour key with Pillow for flat graphics, or pip install rembg and rembg i in.png out.png for photos. Check the edges.
• Check once, lightly. Look at the result (file_view with the image URL works). Confirm the request is met and the required text, names and numbers are right.
• Never generate an image again on your own. Every image costs credits. If a result has a real mistake (a misspelled word, a wrong number, a missing element), show it, say in one sentence what is wrong, and ask whether to make a corrected version. Generate again without asking only when the user told you to keep fixing it until it is right. HTML layout renders cost nothing: fix those yourself.

Deliver

Show generated images in the reply as markdown, ![Short title](url). Sandbox files (renders, processed images) can also be embedded by path, ![Chart](/home/user/out/chart.png). Deliver final files the user should keep with send_attachment. After delivering, stop. Offer one useful next step instead of making extra variants nobody asked for.